Skip to content
The Test Cabinet The Test Cabinet

Overview

Anthropic Claude Code (slug claude) is Anthropic’s command line coding agent, run non-interactively against the test case prompt. It talks to the Anthropic API directly with a user-supplied key, so its runs are billed and served by a single provider rather than routed through an aggregator. See claude.com/claude-code.

Model IDs

Section titled “Model IDs”

Claude Code is driven with vendor-native Anthropic model identifiers — the same names the Anthropic API accepts — passed straight through to the CLI’s --model flag. The following are illustrative, not an exhaustive list:

  • claude-opus-4-8
  • claude-sonnet-4-6
  • claude-haiku-4-5

Invocation

Section titled “Invocation”

The harness probes and invokes the claude binary. Because a harness is installed into the run container at run time rather than baked into an image, each run picks up the most recently published version. The installer runs as the unprivileged run user and drops the binary into ~/.local/bin, which is already on PATH:

Terminal window
curl -fsSL https://claude.ai/install.sh | bash

Each non-interactive session runs the binary with these flags:

claude --print \
  --permission-mode bypassPermissions \
  --output-format stream-json \
  --verbose \
  --model <model> \
  <prompt>

--print runs a single non-interactive turn; --permission-mode bypassPermissions lets the agent act without approval prompts; and --output-format stream-json --verbose selects the line-delimited JSON event stream the harness layer parses (see Events).

Claude Code authenticates with either an Anthropic API key or a Claude account subscription; by default a subscription is preferred when you are signed in. See Authentication for the variables and credential files each mode uses and how to lock the mode.

Because Claude Code drives one provider at one price, the orchestrator records the exact charge the CLI reports rather than looking up per-token prices. Its pricing model is a passthrough — the native model ID is used unchanged — and the cost it reports on its terminal result serves as both the comparable and the actual cost (see Metrics).

See Authentication for the API-key and subscription modes, Events for how Claude Code’s output maps to normalized harness events, Metrics for how its usage and cost are recorded, and the agent harness layer for the contracts both implement.